feat(prune): experiment with time-based retention and gradual catch-up

[meyer9]

Opt-in experiment for the Base Denim 2s → 200ms pruning transition. Replace distance-configured history cutoffs with timestamp-based retention while keeping the existing pruning worker, scheduling, and execution limits. This is not a pruning work/time budget.

• Start each segment at its actual checkpoint; advance the cutoff by 1.3 historical seconds per second of chain progress, capped at head time minus 31 days.
• Persist anchors with prune checkpoints so restarts and snapshot copies preserve progress. Policy changes re-anchor rather than advancing retroactively.
• Preserve ExEx and minimum-distance constraints; prevent backfill writes from discarding data using the superseded distance; check actual pruning progress before unwinds.
• Defaults remain unchanged. Only distance-configured bodies, receipts, account history, and storage history use the opt-in policy.

Enable in reth.toml, alongside the existing distance modes:

[prune.segments.time_based]
retention_seconds = 2678400
catch_up_percent = 130

Snapshot trial: based on Base’s pinned base-v2.5.2.7 release branch. Testing in Base requires rebuilding with this Reth branch and enabling the config; this PR does not update Base’s dependency pin, publish binaries, or modify snapshots. Snapshot consumers need a compatible binary and the enabled config as well as database metadata/checkpoints. Seven days of excess history takes about 23 days to eliminate at 130%. Historical-time advancement is not CPU usage, and fast catch-up can still produce long prune runs. Removing the policy restores block-distance pruning and can cause a catch-up cliff.

Validation: 222 pruning/config/stages/type tests passed (one existing ignored), plus the receipt-write regression passed across storage V1/V2. The real-database transition test advances 50 new 200ms blocks while pruning six old 2s blocks per history segment, survives restarts, and converges. Node/CLI all-target compilation, targeted all-target Clippy with -D warnings, and pinned nightly formatting passed. Workspace all-features Clippy is blocked by missing LLVM 22.1. Snapshot throughput and persistence-backpressure behavior are not yet benchmarked.

Test command:

cargo test --locked -p reth-prune -p reth-prune-types -p reth-config -p reth-stages --lib --features reth-prune-types/reth-codec,reth-config/serde,reth-stages/test-utils
cargo test --locked -p reth-provider --lib test_prunable_receipts_logic

Implementation discussion